python - 为什么对张量的多次调用每次使用 reshape 都会更改其值

Question

我试图在 Jupyter Notebook 中使用 Tensorflow 调试一些代码，并注意到了我没有预料到的行为。我注意到下面所有打印语句的输出都会为 a_S 生成不同的输出矩阵。为什么会这样？我本以为 a_S = tf.random... 行会被计算一次。然后 reshape 运行。由于 reshape 已经发生，第二次和第三次调用不会改变吗？另外，该行中没有任何内容会导致 a_S 更改值，除非它在每次打印时以某种方式重新分配？

with tf.Session() as test:
    tf.set_random_seed(1)
    a_S = tf.random_normal([1, 4, 7, 3], mean=1, stddev=4)
    a_G = tf.random_normal([1, 4, 7, 3], mean=1, stddev=4)
    J_style_layer = compute_layer_style_cost(a_S, a_G)

    m, n_H, n_W, n_C = a_G.get_shape().as_list()    

    #print("J_style_layer = " + str(J_style_layer.eval()))    
    print("J_style_layer2 = " + str(tf.reshape(a_S,[-1,n_C]).eval()))    
    print("J_style_layer3 = " + str(tf.reshape(a_S,[-1,n_C]).eval()))   
    print("J_style_layer3 = " + str(tf.reshape(a_S,[-1,n_C]).eval()))

输出:

J_style_layer2 = [[ -1.68344498e+00   1.89428568e+00   4.18909216e+00]
 [  1.38468242e+00   3.89259148e+00   2.35248661e+00]
 [ -1.92024541e+00   4.64613724e+00  -1.03752756e+00]
 [  4.89945602e+00  -7.53608036e+00   3.40916491e+00]
 [ -2.48588562e-01  -2.43247509e+00   8.14658546e+00]
 [ -1.76469064e+00   2.46943331e+00  -6.58598471e+00]
 [  7.14989901e-01   6.61043763e-01   4.19998550e+00]
 [  2.50623369e+00   4.50162363e+00   1.56363678e+00]
 [  3.42334247e+00  -8.45517635e+00   3.44382668e+00]
 [  1.22528923e+00   2.32022464e-01   1.46155047e+00]
 [  4.21129417e+00  -4.74298620e+00  -2.33842206e+00]
 [  1.37766552e+00  -3.39658976e+00   2.58650005e-01]
 [  1.25126350e+00   7.47679806e+00   5.12217665e+00]
 [  1.06175241e+01  -1.33930969e+00   9.90524054e-01]
 [ -5.41032219e+00  -5.45012522e+00  -1.77322006e+00]
 [ -4.09782457e+00   4.84395790e+00   2.74499822e+00]
 [ -9.30834293e-01   4.23307610e+00   5.83413506e+00]
 [ -2.86698842e+00  -5.36654806e+00   5.01119232e+00]
 [  2.98037910e+00   4.92405033e+00  -5.25441885e-01]
 [ -2.72941589e-03   6.32100248e+00   1.18444359e+00]
 [  2.83025682e-01  -4.18499231e+00   1.21345019e+00]
 [ -2.40400195e+00  -9.58013535e-01   7.72542524e+00]
 [  7.56009007e+00   5.49445009e+00   1.37617004e+00]
 [  5.44784927e+00   5.19741154e+00   4.22853470e+00]
 [  6.33485937e+00   1.05327117e+00   6.67146921e+00]
 [ -7.39221454e-01   3.46544337e+00  -1.65367699e+00]
 [  2.98967314e+00   1.79384637e+00  -1.96472073e+00]
 [  4.62664247e-01  -2.60094047e+00   4.77080643e-01]]

J_style_layer3 = [[ -0.96795654   3.94465876  -3.58533096]
 [  2.86657572   2.92028189   0.78808147]
 [  5.67818737   7.19356251   1.75864506]
 [ -2.02141023   0.41648489  -3.99681711]
 [  6.6851511   -4.29682779  -3.34400749]
 [  4.41373968  -0.86847067   5.63204002]
 [ -9.63679504   8.1419487    3.75494003]
 [ -0.99762154   2.573452     4.2510972 ]
 [ -1.79123497  -4.25948954   0.93429548]
 [ -3.26150131  -1.08424997  -2.17739081]
 [  5.18679285   5.10724735  -1.09539294]
 [  4.88294411  -3.20969486   8.78791428]
 [ -0.57771873  -2.99525881   2.32472777]
 [  1.36514819   2.2696619    2.3027463 ]
 [ -0.12692893   5.43655586   3.8288033 ]
 [  8.78754711   1.60508752  -0.90078032]
 [  0.03909284  -1.96740341   3.91209292]
 [ -2.23007846  -4.52376413   5.86214685]
 [ -2.71961284   1.87416935  10.11991024]
 [  2.14405489   1.23315167  -0.92923737]
 [ -4.21774721   1.86219811   0.18963808]
 [  5.43474102   4.82147026   0.60951805]
 [  5.42611742   1.51715541   4.34838676]
 [  1.46284723   6.85462999   3.54845214]
 [  5.46560287   2.08948946  -4.51927328]
 [  4.75431013   3.07936954   1.04333222]
 [  0.33259553  -0.8477515   -5.67703581]
 [ -0.64754236  -1.51528716   1.61516106]]

J_style_layer3 = [[  3.02279925   3.8460381   -3.10602283]
 [ -2.568856    -3.49679995   4.63807583]
 [  0.43675953   3.30313778  -8.33617496]
 [ -2.65553975   3.8253727    6.25587273]
 [  0.36330599   3.29871988   0.04714996]
 [  1.96590877  -1.24606895  -1.1522367 ]
 [  7.65323305  -2.6084044    0.94568378]
 [ -6.0451417    0.61521298  -3.07440042]
 [  4.33242702   1.00051153   1.34616411]
 [  3.1335001    1.86272466  10.08991051]
 [  0.5303157    0.73839551   7.40760946]
 [ -3.67107534   5.532269    -0.37785971]
 [ -0.60100091  -6.75784874   5.94953918]
 [  4.20541143  -1.72344565  -1.09141064]
 [ -5.15075636  -0.99294972   0.41196495]
 [ -2.43932915   2.13546133   1.18984747]
 [ -0.13836217   1.94724464   4.87366867]
 [ -2.72181702   3.02790046 -13.79762745]
 [ -1.3500247    3.25204515  -3.60114098]
 [  8.59435654  -0.15901434  -0.41931808]
 [  2.08263826  -1.55667186   3.37390661]
 [  2.86219215  -5.58194542  -9.734231  ]
 [  7.96397781   2.00432396  -6.02174616]
 [ -5.0965724    2.81370831  -0.9835161 ]
 [  1.61186719   4.38890743   6.93289137]
 [  4.34721375   0.67843139   4.64758778]
 [ -1.49379516   4.09007645   2.99832821]
 [ -2.06486607   2.20908523   3.9281044 ]]```

Answer 1

当你写

a_S = tf.random_normal([1, 4, 7, 3], mean=1, stddev=4)

您正在做的是在图中创建一个节点， tensorflow 使用该节点来执行您告诉它的任何计算。在这种情况下，您创建的节点将生成随机张量。这与 reshape 功能没有任何关系。您已经创建了一个节点，在评估时会返回一个新的随机张量。您可以看到一个最小的示例，如下所示:

import tensorflow as tf


with tf.Session() as test:
    tf.set_random_seed(1)
    a_S = tf.random_normal([1])
    print("a " + str(a_S.eval()))
    print("b " + str(a_S.eval()))

打印内容

a [-0.67086124]
b [0.9259123]

编辑例如，如果你用它做一些更复杂的事情

import tensorflow as tf


with tf.Session() as test:
    tf.set_random_seed(1)
    a_S = tf.random_normal([1])

    # print statement triggered whenever we go through node b_S
    a_S = tf.Print(a_S, [], message='We went through a_S', summarize=100)

    b_S = 2*a_S
    b_S = tf.Print(b_S, [], message='We went through b_S', summarize=100)

    a_S.eval()
    print('___________')
    b_S.eval()
    # We went through a_S
    # ___________
    # We went through a_S
    # We went through b_S

那么 TF 将只运行它需要的内容来评估您所要求的内容。评估图表某些部分的唯一内容是 .eval 和 session.run(..)。其他一切都是向图中添加东西，无论是张量还是操作。